Sensitivity Stability Analysis of Physics-Informed Neural Networks in Topology Optimization: PINN vs Gradient-Enhanced PINN
Journal of Algorithms in Complex Systems · 2026
DOI: 10.66879/jacs.2026.56 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Algorithms in Complex Systems, 2026. Verfasst von Haroon Ijaz.
Abstract (Originalsprache)
Physics-informed neural networks (PINNs) have attracted considerable interest as mesh-free, differentiable alternatives to finite element solvers in topology optimization pipelines. However, their practical utility is undermined by a fundamental instability in the sensitivity fields they produce: higher-order spatial derivatives computed via automatic differentiation exhibit high variance and spatial inconsistency, causing gradient-based design updates to diverge and producing irregular optimized layouts. This paper presents a systematic sensitivity stability analysis comparing standard PINNs and gradient-enhanced PINNs (gPINNs) as structural solvers within a SIMP-based topology optimization framework. We evaluate four statistical measures of sensitivity quality — mean, median, standard deviation, and maximum variance — across 50 optimization iterations on three benchmark beam configurations (cantilever, simply supported, and fixed–fixed). Results demonstrate that the gPINN formulation achieves a 94.35\% reduction in mean sensitivity variance relative to standard PINNs, with correspondingly smoother optimized topologies consistent with classical finite element benchmarks. A post-filtering analysis further shows that gPINN requires substantially less external smoothing to achieve stable optimization convergence, confirming that gradient enhancement provides intrinsic rather than post-hoc stability. These findings establish sensitivity stability as a critical and previously underexamined criterion for evaluating physics-informed surrogates in structural optimization.
Zitieren
APA
Ijaz, H. (2026). Sensitivity Stability Analysis of Physics-Informed Neural Networks in Topology Optimization: PINN vs Gradient-Enhanced PINN. Journal of Algorithms in Complex Systems. https://doi.org/10.66879/jacs.2026.56
Vancouver
Ijaz H. Sensitivity Stability Analysis of Physics-Informed Neural Networks in Topology Optimization: PINN vs Gradient-Enhanced PINN. Journal of Algorithms in Complex Systems. 2026. doi:10.66879/jacs.2026.56
BibTeX
@article{Ijaz2026,
author = {Ijaz, Haroon},
title = {Sensitivity Stability Analysis of Physics-Informed Neural Networks in Topology Optimization: PINN vs Gradient-Enhanced PINN},
journal = {Journal of Algorithms in Complex Systems},
year = {2026},
doi = {10.66879/jacs.2026.56}
}
EUVIMED